💓 PulsePod

A Robust Framework for Arrhythmia Detection using Deep and Spiking Neural Networks

Deep Learning • Neuromorphic Computing • Real-Time Deployment

Overview • Tech Specs • Architecture • Installation • Results

---

🩺 Project Overview

PulsePod is an end-to-end AI system designed to detect cardiac abnormalities from ECG signals with high reliability and efficiency. Unlike traditional models that fail when deployed on new patient populations (the Dataset Shift problem), PulsePod is built for generalization.

By training on a massive, unified dataset combining MIT-BIH and PTB-XL records , the system achieves robust performance across diverse demographics. It features a suite of models ranging from high-accuracy 1D-CNNs to ultra-low-power Spiking Neural Networks (SNNs) designed for battery-constrained wearable devices

🌟 Key Features

• 🛡️ Robust Generalization: Achieved 87% Accuracy on the held-out Unified Test Set (61,126 samples), successfully generalizing across diverse signal sources.
• 🧠 Neuromorphic Intelligence: Validated SNNs (Leaky Integrate-and-Fire) achieving 85% accuracy on the unified dataset, demonstrating viability for energy-efficient edge hardware.
• ⚖️ Stratified Training: Models trained on 244,502 heartbeats with balanced class distribution to prevent bias and overfitting.
• 🚀 Real-Time Dashboard: Interactive Streamlit interface for visualizing live ECG inference and probability confidence.

---

🔬 Technical Specifications

The system was engineered and validated using rigorous quantitative benchmarks.

📊 Data Composition

• Total Processed Heartbeats: 305,628 (Train + Test)
• Source Integration:
  • MIT-BIH: 54,680 beats (Gold-standard arrhythmia annotations)
  • PTB-XL: 250,948 beats (Diverse 12-lead clinical data)
• Preprocessing: NeuroKit2 pipeline with 187-sample fixed window segmentation

🧠 Model Architectures

1. Deep CNN (Robust Classifier)

• Structure: 2-Layer 1D-Convolutional Network with MaxPooling and Dropout (0.5)
• Kernels: 32 & 64 filters (Kernel Size=5) optimized for morphological feature extraction
• Loss Function: Focal Loss implemented to counter class imbalance

2. Spiking Neural Network (Neuromorphic)

• Neuron Model: Leaky Integrate-and-Fire (LIF) with surrogate gradient descent
• Simulation: 50 Time Steps per inference window:
• Decay Rate: $\beta = 0.95$ (Optimized for membrane potential retention)

---

🏗 Architecture

The system follows a modular pipeline from raw signal ingestion to edge deployment.

graph LR
    A[Raw ECG Data] --> B(Preprocessing & Segmentation)
    B --> C{Unified Dataset Creation}
    C --> D[Training Loop]
    D --> E[Model Zoo]
    E --> F[Deployment Dashboard]

    subgraph Models
        E1[Standalone CNN]
        E2[SNN Neuromorphic]
        E3[CNN+LSTM]
    end

    E --> E1
    E --> E2
    E --> E3

Loading

---

⚙️ Installation

To ensure environment stability, requirements are split into two categories.

1. Clone the Repository

git clone https://github.com/Vishwanath-06/PulsePod.git
cd PulsePod
git lfs install
git lfs pull

2. Set up the Environment

Option A: For Standard CNN Models (Recommended) Use this for the main Dashboard and standard Deep Learning models.

# Create virtual environment
python -m venv venv_cnn
source venv_cnn/bin/activate  # or venv_cnn\Scripts\activate on Windows

# Install dependencies
pip install -r cnn_requirements.txt

Option B: For Neuromorphic SNN Experiments Use this specifically for running Spiking Neural Networks.

# Create virtual environment
python -m venv venv_snn
source venv_snn/bin/activate  # or venv_snn\Scripts\activate on Windows

# Install dependencies
pip install -r snn_requirements.txt

---

🚀 Usage

1. Launch the Dashboard

Visualize the inference stream in a web interface.

# Ensure venv_cnn is active
streamlit run app.py

2. Data Preprocessing (Optional)

The test data is already provided in .npy format in the data/ folder. If you wish to regenerate the unified test set from raw sources:

python data/create_unified_test_set.py

---

📊 Performance & Results

We evaluated our models on a held-out, multi-domain test set (Unified MIT-BIH + PTB-XL, 61,126 samples) to ensure true robustness.

Model Architecture	Accuracy	Precision (Abnormal)	Recall (Abnormal)	F1-Score	Use Case
Robust CNN	87%	0.89	0.86	0.88	Best Overall Balance
Spiking NN (SNN)	85%	0.85	0.87	0.86	Ultra-Low Power / High Recall
CNN + LSTM	75%*	0.95	0.58	0.72	Precision / Specificity

> Note: While CNN-LSTM performed well on clean data, the standalone CNN demonstrated superior generalization on the heterogeneous PTB-XL test set (82% vs 75%).

📈 Detailed Evaluation

The confusion matrices below demonstrate the Robust CNN's ability to minimize False Negatives (critical for healthcare), while the metrics comparison highlights the SNN's competitive performance despite being energy-efficient.

Confusion Matrices (Generalization Test)

Model-wise Metrics Comparison

📂 Project Structure

PulsePod/
├── data/                           # Processed .npy datasets & Scripts
│   ├── mit_bih/
│   ├── ptb_xl/
│   ├── unified_test_set/
│   └── create_unified_test_set.py  # Script to merge datasets
├── models/                         # Pre-trained PyTorch models
│   ├── final_robust_cnn.pt
│   ├── final_robust_cnn_focal_loss.pt
│   ├── final_robust_cnn_lstm.pt
│   └── final_robust_snn.pt
├── app.py                          # Main Streamlit Dashboard
├── cnn_requirements.txt            # Standard dependencies
├── snn_requirements.txt            # SNN dependencies
└── readme.md

---

🔮 Future Scope

• Hardware Porting: Deploy SNN models to Intel Loihi or Neuromorphic chips.
• Multi-Class: Extend detection to specific arrhythmias (AFib, PVC, LBBB).
• Federated Learning: Implement privacy-preserving patient data training.

Built with 💙 by Vishwanath.T